Predicting Asthma Attacks: Effects of Indoor PM Concentrations on Peak Expiratory Flow Rates of Asthmatic Children

Dohyeong Kim, Sunghwan Cho, Lakshman Tamil, Dae Jin Song, Sungchul SeoIEEE Access 2020 · DOI

Introduction

Asthma remains a significant public health challenge, particularly for children whose respiratory systems are highly sensitive to environmental triggers. While the correlation between air pollution and asthma exacerbations is well-established, traditional management strategies often rely on broad, retrospective data. This study presents a pilot framework for real-time, personalized asthma risk prediction using Internet of Things (IoT) sensors and deep learning.

The primary "predictive gap" addressed here is the lack of immediate, actionable evidence for caregivers. Indoor air quality can fluctuate significantly within minutes, and its impact on a child's Peak Expiratory Flow Rate (PEFR)—a critical metric for lung function—is cumulative. Traditional linear models often fail to capture these complex, time-lagged relationships, necessitating more sophisticated temporal modeling.

Methodology

The researchers implemented a multi-stage data-driven framework to bridge the gap between environmental monitoring and clinical prediction.

Real-time Environmental Monitoring

Low-cost IoT sensors were deployed in the homes of 14 pediatric patients. These devices recorded PM10 and PM2.5 concentrations, temperature, and humidity at 10-minute intervals over a full year (September 2017 to August 2018).

Patient Clustering

To account for individual variability in asthma severity, the study employed K-means clustering. Patients were divided into two distinct groups based on their PEFR stability. Cluster 2 represented a more vulnerable cohort, characterized by lower average lung function and higher symptom variability.

Comparison of average Peak Expiratory Flow Rates (PEFR) between the two patient clusters over the study period.

Deep Learning Prediction

The core innovation of this research is the application of Long Short-Term Memory (LSTM) networks to forecast asthma risk. Unlike standard Multinomial Logistic (MNL) regression, LSTMs are designed to "remember" historical data, making them uniquely suited for medical forecasting where past exposure influences current health status.

The model was trained to classify risk into three categories: Green (Normal), Yellow (Mild/Moderate exacerbation), and Red (Severe exacerbation).

LSTM model architecture for predicting PEFR severity levels based on historical patient and environmental data.

Key Insights and Results

The study's findings highlight the superiority of temporal deep learning over traditional statistical methods.

The "Memory" Advantage

The LSTM model significantly outperformed MNL regression, showing a 25-27% increase in precision for Cluster 1 and a substantial 57-84% increase for the more severe Cluster 2. This performance boost is attributed to the LSTM's ability to process the cumulative effects of particulate matter. Sensitivity analysis further revealed that PM measurements from the hour immediately preceding a PEFR recording were the most potent predictors of decline.

Targeted Effectiveness

Interestingly, the model demonstrated higher accuracy for the more vulnerable patient group (Cluster 2). While traditional models struggled to identify the "Red" (severe) zone, the LSTM successfully captured these high-risk events. This suggests that AI-driven monitoring is most effective for those at the highest risk of life-threatening attacks.

Conclusion and Future Directions

This pilot study confirms the feasibility of using IoT and deep learning for real-time asthma management. By treating asthma risk as a temporal sequence problem rather than a series of snapshots, the researchers achieved a new level of predictive precision.

However, as a pilot study with 14 participants, several limitations remain. Future research should expand the cohort size and incorporate additional variables such as medication adherence, physical activity, and outdoor air quality data. Despite these hurdles, the work paves the way for "smart" asthma management systems that can provide parents with the early warnings needed to prevent severe attacks.

Summary updated

BibTeX
@article{kim2020predicting,
  author = {Dohyeong Kim and Sunghwan Cho and Lakshman Tamil and Dae Jin Song and Sungchul Seo},
  title = {{Predicting Asthma Attacks: Effects of Indoor PM Concentrations on Peak Expiratory Flow Rates of Asthmatic Children}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2019.2960551},
}
Figure